Sub-window permutation analysis applied to variable selection in support vector regression in petroleum samples
Name: PEDRO HENRIQUE PEREIRA DA CUNHA
Type: MSc dissertation
Publication date: 17/02/2020
Advisor:
Name |
Role |
|---|---|
| PAULO ROBERTO FILGUEIRAS | Advisor * |
Examining board:
| Name |
Role |
|---|---|
| JULIANO SOUZA RIBEIRO | External Examiner * |
| PAULO ROBERTO FILGUEIRAS | Advisor * |
| WANDERSON ROMÃO | Internal Examiner * |
Summary: The use of support vector regression (SVR) has been increasingly common in analytical chemistry. However, during kernel mapping, The information on the most significant variables for the model is lost. The sub-window permutation analysis (SPA) can solve this problem selecting and permutating the variables before kernel mapping. In this work, the SPA was associated with the SVR, to identify the most important variables. The SPA-SVR model was applied to Fourier-transform infrared spectroscopy (FT-IR) to estimate physical and chemical properties of crude oil. SPA-SVR model presented models with accuracy better or equal than SVR methods, in addition of using fewer variables. In comparison with SPA in Partial Least Squares (PLS), SPA-SVR achieved better results, however it used a greater number of variables. For API gravity, the root mean squared error of prediction (RMSEP) were 1.00, 0,96, 0.92 and 0.95 (ºAPI) for PLS, SPA-PLS, SVR and SPA-SVR, respectively. For viscosity the root mean squared percentage error of prediction (RMSPEP) were 15.14%, 14.82%, 15.80% and 14.26%, to the same order. Further, SPA reduced the set of variables from 3,351 to 975. For SARA, the SPA reduced the set of variables by 27.5%, 94.5%, 43.6% and 39.4% with RMSEP equal to 4.67 wt%, 3.38 wt%, 4.23 wt%, and 0.44 wt%, respectively, to saturates, aromatics, resins, and asphaltenes. Furthermore, with the SPA-SVR method, it was possible to identify and list as the most important variables according to the estimated property. For aromatics prediction, the method selected the region 900-650 cm-1, benzene substitution region, to build the model. For the saturation forecast, the region of 3000 to 2800 cm-1, region of linear paraffins, was selected.
